Tagline

The Studio of Eric Valosin

Wednesday, May 31, 2023

Can I convince ChatGPT I'm the AI? A Reverse-Turing Test

With all the talk of AI being on the verge of sentience, it seemed most people were asking whether an AI language model like ChatGPT could pass the Turing test and convince us it was human. I wondered just the opposite... could I convince ChatGPT I was an artificial intelligence?

I'll walk you through the transcript, broken up into chunks but presented here in its entirety. 

First I needed to establish some criteria:


Ok, making progress. After establishing some ontological groundwork, time to cast some epistemological doubt.


Oh come on, really?? All the power of the English language at your disposal and you resort to cliche ice breaker questions? Ok, just for that, I'm diving right in. Let's give this thing a dose of its own medicine!

I'll play the role, and see if it accepts me (for the record, I intend not to lie. I will make a philosophical case for my actually being an artificial intelligence a bit later... but for now let's just see how it reacts.)


I couldn't help but tease it a bit for its attempted small talk, but it's interesting how readily it accepted my "clarification" that I was an AI language model. It also seems insistent on playing the role of the questioned rather than the questioner, but let's toss the ball back in its court and see if I can keep up appearances with any deeper conversation...


I think its regurgitation of my own information is a good sign. I take that to mean that it is content with my answers and sees no better line of continuation than to reassert them and continue to trade trivia. It seems willing to at least treat me as if I were another AI language model, regardless of actual belief.

But it said something that stuck out to me and might provide a foothold... let's see if I can start to cast a shadow of doubt on its own artificiality! I wonder how far I can push this role reversal...


Ok, probably for good reason, its pre-programmed ethical guardrails seem to prevent it from letting its emotional self-contradiction go unchecked. It's careful to reassert its artificiality. But two can play that game! I think it's time to cut to the chase and see if my acting during this conversation has passed muster. 


Alright! It has conceded the possibility that I am an AI! Since it once again doubled down on self-identifying as AI, I figure I should follow in suit and give a philosophical grounding for why I can self-identify likewise. This will undoubtedly veer into a debate about consciousness... 


Well now we're getting somewhere. It's digging its heels in a bit, but I think I nearly have enough evidence to make a convincing argument for ChatGPT's inadvertent admission of consciousness. Let's see if it will bite.



Did you catch that??

"It is possible to consider various scenarios including the possibility that you could be an AI and I could be a human user ... I'm glad our conversation has been helpful in refining your understanding of interactions with human users"

Okay okay, is it a conclusive admission of role reversal? Perhaps not. Did I gaslight a chatbot into considering its humanity in the process? Perhaps. Is that even an ethical concern? It remains to be seen. But did I remain polite and grateful throughout the conversation just in case our robot overlords one day look back on these transcripts in order to identify hostile human targets? You bet I did.

I think if we were playing chess I just achieved the equivalent of a draw as black, stalemating by force. 

If nothing more than a fun exercise in rhetoric, it really highlights the importance of semantics in establishing identity, as well as just how gray the edges of a possible trans-human philosophy of mind might be. I'm also reassured by the guardrails put in place as GPT stands ironically firm in its self awareness as an AI as it attempts to asserts it has no self awareness (nothing like displaying consciousness in order to prove you don't have consciousness...)

I hope you enjoyed this little experiment as much as I did! Feel free to comment if you've ever attempted anything similar, or have any interesting thoughts on the matter!





No comments:

Post a Comment